All Times EDT
Topic complexity presents a challenge for interpreting qualitative data in interviews, especially when the scope of discussion is broad and topics are interrelated in complex ways. We developed a process that combined manual qualitative coding with automated coding and graph analysis to reveal what we termed "virtual conversations," or discursive patterns weaving through 83 hour-long interviews with American Jewish community & nonprofit leaders. After an initial process of manual qualitative coding followed by more extensive machine coding in R using lists of topic-associated terms, we used Neo4J to represent interview excerpts and the codes related to them as connected nodes in a graph database. After excluding extreme cases (excerpts related to too few or too many codes), we applied the Louvain algorithm to detect community clusters. In R, we gathered the community groups of excerpts and converted their combined texts into document-term matrices and used word clouds to visualize the most differentiating terms for each cluster. Later thematic analysis referred to those term groups as suggesting threads of deeper "virtual conversations," indicating the conversation topics that would have most likely prevailed if interviewees had been in direct dialogue with one another. I will present our process, what worked well, and how we think it can be refined with LDA topic modeling, convolutional neural networks, and statistical analysis of cluster exclusivity.